Re: [Qemu-devel] [RFC v2 15/22] monitor: send event when request queue full

[Stefan Hajnoczi]

On Thu, Oct 19, 2017 at 03:16:11PM +0800, Peter Xu wrote:
> On Wed, Oct 18, 2017 at 05:28:04PM +0200, Stefan Hajnoczi wrote:
> > On Mon, Oct 16, 2017 at 04:11:58PM +0800, Peter Xu wrote:
> > > On Thu, Oct 12, 2017 at 01:56:20PM +0100, Stefan Hajnoczi wrote:
> > > > On Fri, Sep 29, 2017 at 11:38:37AM +0800, Peter Xu wrote:
> > > > > Set maximum QMP request queue length to 8.  If queue full, instead of
> > > > > queue the command, we directly return a "request-dropped" event, 
> > > > > telling
> > > > > client that specific command is dropped.
> > > > > 
> > > > > Signed-off-by: Peter Xu <address@hidden>
> > > > > ---
> > > > >  monitor.c | 15 +++++++++++++++
> > > > >  1 file changed, 15 insertions(+)
> > > > > 
> > > > > diff --git a/monitor.c b/monitor.c
> > > > > index 1e9a6cb6a5..d9bed31248 100644
> > > > > --- a/monitor.c
> > > > > +++ b/monitor.c
> > > > > @@ -3971,6 +3971,8 @@ static void monitor_qmp_bh_dispatcher(void 
> > > > > *data)
> > > > >      }
> > > > >  }
> > > > >  
> > > > > +#define  QMP_ASYNC_QUEUE_LEN_MAX  (8)
> > > > 
> > > > Why 8?
> > > 
> > > I proposed this in previous discussion and no one objected, so I just
> > > used it. It's here:
> > > 
> > >   https://lists.gnu.org/archive/html/qemu-devel/2017-09/msg03989.html
> > >   (please don't go over the thread; I'll copy the related paragraphs)
> > > 
> > > """
> > >   ...
> > >   Regarding to queue size: I am afraid max_size=1 may not suffice?
> > >   Otherwise a simple batch of:
> > > 
> > >   {"execute": "query-status"} {"execute": "query-status"}
> > > 
> > >   Will trigger the failure.  But I definitely agree it should not be
> > >   something very large.  The total memory will be this:
> > > 
> > >     json limit * queue length limit * monitor count limit
> > >         (X)            (Y)                    (Z)
> > > 
> > >   Now we have (X) already (in form of a few tunables for JSON token
> > >   counts, etc.), we don't have (Z), and we definitely need (Y).
> > > 
> > >   How about we add limits on Y=16 and Z=8?
> > > 
> > >   We can do some math if we want some more exact number though.
> > >   ...
> > > """
> > > 
> > > Oops, I proposed "16", but I used "8"; I hope 8 is good enough, but I
> > > am definitely not sure whether "1" is good.
> > 
> > I understand the concern about breaking existing clients but choosing an
> > arbitrary magic number isn't a correct solution to that problem because
> > existing clients may exceed the magic number!
> 
> I agree.
> 
> > 
> > Instead I think QMP should only look ahead if the out-of-band feature
> > has been negotatiated.  This way existing clients continue to work.  New
> > clients will have to avoid sending a batch of requests or they must
> > handle the queue size limit error.
> 
> Hmm yes I just noticed that although I broadcasted the "OOB"
> capability but actually I skipped the negociation phase (so OOB is
> always enabled). I think I should have that for sure.
> 
> IIUC below new handle_qmp_command() should be always compatible with
> old clients then:
> 
> handle_qmp_command ()
> {
>   ...
>   if (oob_enabled) {
>     if (cmd_is_oob (req)) {
>       // execute command
>       qmp_dispatch (req);
>       return;
>     }
>     if (queue_full (mon)) {
>       // drop req
>       send_full_event (mon);
>       return;
>     }
>   }
> 
>   queue (req);
>   kick (task);
> 
>   if (!oob_enabled) {
>     // if oob not enabled, we don't process next request before previous
>     // one finishes, and queue length will always be either 0 or 1.
>     // Note: this means the parsing thread can block now.
>     wait_until_req_handled (req);
>   }
> }
> 
> This will be somehow more complicated than before though, since if
> with this, we need to make sure all the QMP clients have enabled OOB
> feature to make sure OOB command can work. Otherwise even if only one
> QMP client didn't enable OOB, then it may block at waiting for the
> request to finish, and it will block the whole monitor IOThread as
> well (which is currently shared by OOB and non-OOB monitors).
> 
> Or, maybe, I should just create one IOThread for each QMP monitor.

Or temporarily stop monitoring a client's chardev while the request is
being processed if OOB isn't negotiated.  That way a single IOThread can
still service multiple QMP monitors with differing OOB settings.

Stefan